Physical-informed neural network in modeling and control of a servo mechanical rotor system
Yi-Ho Chen, Chao-Chung Peng · Results in Engineering · 2025
• A hybrid physical-informed data driven modeling method is presented. • Nonlinear model can be precisely modeled by fewer parameters. • Robust nonlinear feedback controller is presented via LMI. • Stability proof via Lyapunov is applicable to the hybrid control system. • An application for a real physical system validates the feasibility. In this paper, a neural network based system identification method is proposed to a class of nonlinear 2nd order mechanical rotor system with unmeasurable internal dynamics. Combined with physics-based and data-driven approaches, the proposed neural network model is merged as physical-informed dynamics, enabling it to effectively capture the unmeasurable internal dynamics. By considering the order of the identified system and physics-informed terms, the proposed method performs a robust and outstanding modeling result not only on the steady-state but also on the transient behavior of the system. The proposed framework also benefits the controller design variety, as the identified model can be formulated by a standard system form, making it possible to apply well-known control theorem to meet the requirement. For application and validation point of view, this work considers a 2nd order electrical-driven industrial cooling fan with significant transient response and compares the proposed algorithm with different commonly used system identification techniques. In addition, a robust PID controller is designed based on the identified physical-informed neural network model. The result demonstrates that the proposed model achieves a superior modeling and control result with fewer parameters compared to the conventional neural network structure, making it suitable for real-world engineering implementation.